Position-Aware Attention Mechanism-Based Bi-graph for Dialogue Relation Extraction

Guiduo Duan, Yunrui Dong, Jiayu Miao, Tianxi Huang

Abstract: Relation extraction in a dialogue scenario aims to extract the relations between entities in a multi-turn dialogue. Unlike the conventional relation extraction task, the dialogue relation cannot yield a result through a single sentence. Therefore, it is essential to model multi-turn dialogue for reasoning. However, dialogue relation extraction easily causes referential ambiguity owing to the low information density in the dialogue dataset and a large amount of pronoun referential information in the dialogue. In addition, most existing models only consider the token-level information interaction and do not fully utilize the interaction between discourses. To address these issues, a graph neural network–based dialogue relation extraction model is proposed using the position-aware refinement mechanism (PAR-DRE) in this paper. Firstly, PAR-DRE models the dependencies between the speaker’s relevant information and various discourse sentences and introduces pronoun reference information to develop the dialogue into a heterogeneous reference dialogue graph. Secondly, a position-aware refinement mechanism is introduced to capture more discriminative features of nodes containing relative location information. On this basis, an entity graph is built by merging the abovementioned nodes, and the path reasoning mechanism is used to infer the relation between entities in the dialogue. The experimental results on the dialogue dataset indicate that the performance of the F1 value of this method is enhanced by 1.25% compared with the current mainstream approaches.